Ø1 Perucho
72.2% to 76.9%
exact recall improvement
IN PRODUCTION, 170 days
- System
- Autonomous WhatsApp agent that answers customers, transcribes voice notes and builds quotes from live POS prices over an 8,000+ SKU hardware catalog.
- Problem
- Customers write in colloquial Spanish, with typos and voice notes, and tradespeople ask for quantities ('how much cement for 400 blocks') that generic LLMs get wrong.
- n8n
- WAHA
- Docker
- Node.js 22
- OpenRouter / OpenAI
- text-embedding-3-small
- Groq Whisper
- Supabase PostgreSQL (pg_trgm, pgvector, RLS, Realtime)
- TypeSafe Jev
- Vercel AI Gateway
- Engram
- Next.js 14
- Tailwind CSS
- Vitest
- Playwright
- Result
- Built an n8n workflow with WAHA on Docker, a 5-layer retrieval cascade with pg_trgm and pgvector HNSW indexing, Groq Whisper transcription, Supabase Realtime escalation queues, and a Next.js 14 Flight Deck CRM. Raised exact recall from 72.2% to 76.9% across a 320-case eval bench and served 191 customers.
Started Jun 5, 2026
As of 2026-11-22





